acceptodds
Withdrawn from ICLR 2027

LatentFlow: Componentwise Nonlinear Continuation for Multivariate Time-Series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multivariate forecasting architectures increasingly expose structure through decomposition, temporal scales, spectral routing, or variable-aware representations. Yet a distinct design question remains less directly isolated: once several structured temporal coordinates are available, when should their coordinate axis be reduced relative to the nonlinear forecasting operator? We study this representation–continuation interface with LATENTFLOW, an architecture for componentwise nonlinear continuation. A prefix-only encoder constructs structured coordinates, and shared coordinate-modulated variate–patch operators transform each coordinate separately through horizon projection and decoding; only the resulting future forecasts are then aggregated. A complementary linear bank handles simple extrapolative behavior. Under a controlled fixed-budget protocol with 2,048 matched training windows per task, LATENTFLOW achieves the lowest aggregate MSE and MAE among seven evaluated implementations, wins 23 of 28 seed-matched MSE comparisons against TimePro, and retains a 1.38% mean task-wise MSE advantage across five seeds. More importantly, a matched intervention that reduces the same coordinate axis before nonlinear continuation increases mean relative MSE by 5.12% (95% source–seed clustered interval [2.42, 8.44]%). Conversely, replacing the GP-derived coordinates with a matched deterministic multibranch representation yields nearly identical aggregate point-forecasting error and an even 14–14 task split. Together, these results show that the main effect is not tied to a particular stochastic parameterization, but to the computation that consumes the representation: structured coordinates can be more effective when they remain computationally explicit through nonlinear continuation and are fused only after their futures are formed.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, multivariate time series forecasting, machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_7diDOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies multivariate long-term forecasting. It splits each series into a few components (periodic, smooth, short-memory) with a sparse Gaussian process, applies the same nonlinear block to each component separately, and sums only the resulting forecasts ("late fusion"). The output is then mixed with a linear bank and with a "frozen reference" predictor freff_{\mathrm{ref}}, and validation MSE selects between the mixture and freff_{\mathrm{ref}} alone (Eq. 13). Under a reduced protocol (2,048 training windows per task, all baselines rerun by the authors on 7 datasets and 4 horizons), the paper reports the lowest mean MSE, a 1.38% gain over TimePro across five seeds, and a 5.12% degradation when the components are summed before the nonlinear block instead of after (Table 2). A deterministic replacement of the GP front end gives the same accuracy (Table 14). The theory (Lemma 1, Proposition 1) only states that the two fusion orders are different computations and does not say which is better (Section 3.4).

Critical strengths

S1. The authors report results that go against their own method and do not hide them: Table 14 (a deterministic front end is as good as the GP front end), Table 16 (a single resolution is as good as multi-resolution), and Table 21 (channel-order sensitivity). They also use five seeds, a 2×2×22\times2\times2 factorial design (Tables 17--18), and report two kinds of confidence intervals in Section 5.

S2. The question itself, whether decomposed components should be summed before or after a shared nonlinear forecaster, is a reasonable question for a controlled comparison.

Critical weaknesses

W1. The main result cannot be attributed to the described method because of an undefined component, freff_{\mathrm{ref}}. The paper never specifies what model freff_{\mathrm{ref}} is; only its training schedule is given (Appendix A.4). It is a candidate of the final validation selection (Eq. 13), and freff_{\mathrm{ref}} alone reaches MSE 0.35138 (Section 5.2), which equals the strongest baseline TimePro (0.35136, Table 10). Everything described in Section 3 adds only 1.22% on top of it.

W2. The baseline comparison is asymmetric. Appendix D says that recommended, dataset-specific configurations were used, but then gives one reduced setting for iTransformer, TimeMixer and VPNet across all datasets (width 128 and two layers; one downsampling level; a fixed 3x3 kernel), each trained for 10 epochs (Table 9). LatentFlow is trained in three stages (10 + 2 + 10 epochs) with per-dataset settings selected by validation (Table 5). A unified reduced setting can be reasonable, but then it should also apply to the proposed model, or the paper should show that the official per-dataset settings do not perform better under the 2,048-window protocol; neither is done. No code or supplementary material is provided.

W3. Several terms are never defined ("prototype-memory operator", "scale head", freff_{\mathrm{ref}}), and the symbols in Eqs. 9-13 are not defined where they appear. Table 2, the only mechanism table in the main text, is never cited, and Tables 4, 6, 7, 8, 9, 11, 21 and 23 are not cited anywhere either, so the reader is not told what these tables are meant to show. The layout is unpolished: Table 3 appears right after the references, before the Appendix A heading. The reproducibility statement promises per-horizon results in the appendix, but none are given. ETSformer is cited as ICLR 2023 but was rejected there and exists only as arXiv:2202.01381.

Questions for authors

Q1. What is freff_{\mathrm{ref}}? Please state which model it is (its layers, inputs, and training data) and how it relates to the baselines in Table 1. Please also report Tables 1 and 2 with freff_{\mathrm{ref}} removed from the candidate set in Eq. 13, so that the contribution of the method described in Section 3 can be seen on its own.

Q2. Why were the baselines run with a single reduced setting rather than their official per-dataset settings? If official settings were tried under the 2,048-window protocol, please report them, at least for Electricity and Traffic where the margins are largest.

Q3. The reproducibility statement says the appendices give per-horizon results, but all appendix tables are 4-horizon averages or the 720 horizon only. Where are the per-horizon results? Also, what do "prototype-memory operator" and "scale head" refer to?

Reviewer_BfEcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes LatentFlow, a multivariate time-series forecasting framework that investigates when structured temporal components should be aggregated relative to nonlinear forecasting operations. The central idea is to preserve individual latent components throughout nonlinear continuation and horizon decoding, and aggregate them only in the forecast space, rather than fusing them before nonlinear processing. The authors conduct experiments on seven datasets with 28 forecasting tasks under a fixed-budget setting. The results show that LatentFlow achieves competitive forecasting accuracy and outperforms an early-fusion variant by 5.12% in mean relative MSE.

Critical strengths
  • The authors provide several targeted interventions, including early-fusion comparisons, deterministic multibranch representations, linear-bank ablations, and factorial analysis. These experiments go beyond conventional module-removal ablations and provide useful evidence regarding the proposed design. I like these experiments design.

  • The paper includes multiple random seeds, paired comparisons, clustered bootstrap confidence intervals, and checkpoint-selection audits. These analyses improve the credibility of the reported results.

Critical weaknesses
  • Although LatentFlow achieves the best average MSE among the evaluated methods, its five-seed mean relative improvement over TimePro is only 1.38%. Meanwhile, the proposed architecture involves multiple coordinate paths, nonlinear continuation operators, and additional representation-learning components. The reported computational analysis also indicates substantially higher inference costs than lightweight alternatives such as VPNet. The paper would benefit from a more systematic accuracy–efficiency analysis, particularly under matched computational budgets rather than only matched numbers of training windows.

  • The main experiments are conducted with 2,048 training windows per task and a fixed input length of 96. While this controlled setting is reasonable for investigating architectural inductive biases, it remains unclear whether the reported advantages generalize to standard full-data forecasting settings. In particular, the improvements may partly arise from the stronger structural inductive bias under limited supervision. Experiments with different training-data budgets and longer input sequences would substantially strengthen the generality of the conclusions.

  • The paper demonstrates that preserving structured components can improve performance, but it provides limited insight into the conditions under which this advantage emerges. For example, Proposition 1 relates the discrepancy between early and late fusion to nonlinear curvature and dispersion among components. However, the experiments do not directly examine whether component diversity, correlation, or other representation characteristics are associated with forecasting improvements. Such analyses would provide a more informative explanation of when and why componentwise continuation is useful, rather than merely demonstrating that it works in the evaluated architecture.

Questions for authors
  1. How can the authors distinguish the benefit of late fusion from the benefit of maintaining diverse branch inputs and coordinate-specific nonlinear activations? Would a capacity-matched single-stream early-fusion baseline lead to similar conclusions?

  2. Can the authors provide additional analysis connecting the theoretical discrepancy between early and late fusion to actual forecasting improvements? For example, is the improvement correlated with component dispersion, component correlation, or nonlinear operator characteristics?

  3. How sensitive are the results to different look-back lengths?

  4. Given the relatively modest improvement over TimePro, how does LatentFlow compare with existing methods under matched training-compute or inference-compute budgets?

Reviewer_CcJ6OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The work first identifies that many time series forecasting models use explicit inductive biases on extracted features, such as finding trend decompositions or explicit periodicity features. The work investigates whether such intermediate vectors should be forecast individually and then fused, or fused first and then forecast. The method shows that component-wise forecasts and late fusion are beneficial. This is supported by empirical evidence and a proof showing a qualitative difference between the orders.

Critical strengths
  1. The question of when to fuse certain representations (e.g., before or after forecasting/early vs. late fusion) is relevant. It might unlock gains in accuracy and improve interpretability (the latter is not explored in the work).
Critical weaknesses
  1. Poor clarity: The paper does not follow typical conventions, making it hard to follow or parse in a reasonable time. For instance (non-exhaustive):
    • The intro contains not a single reference to the extensive existing literature.
    • The introduction already provides methodological details, theoretical results, and empirical measurements.
    • Figures are too small to read (Fig. 1) or have broken layout (Fig. 2).
    • App. I and J are briefly mentioned at the very end, but no summary or findings from them are provided. App. G and H are never even mentioned in the main text.
    • Standard notation (vectors and matrices are bold) is not followed.
  2. The empirical analysis is not convincing.
    • The investigated datasets are limited. The authors should consider a broader evaluation beyond/instead of the datasets investigated, which are somewhat flawed and biased. For a start, GIFT-Eval and fev-bench could be considered.
    • The reported baselines fall far short of what has been published by the respective original works (e.g., iTransformer). The "matched reruns" (ll. 366f) don't match the previous results.
    • The baselines are dated and don't cover the current progress of the field. The work should mention and compare against models such as TimeMixer++, TiDE, Chimera, xLSTM-Mixer, and potentially newer ones.
    • The evaluation protocol differs from the typical one in the literature. See Wu et al. (2021) and Zhou et al. (2021).
    • Existing, established baselines outperform the shown models, including LatentFlow, substantially. See Kraus et al. (2025) for an overview.
    • No implementation is provided, severely limiting reproducibility.

Note: Both weaknesses are too extensive to be fixed in a rebuttal. A majorly reworked version of this paper, however, might be a valuable contribution.

References:

  • Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long (2021). Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting. In Proceedings of the Conference on Neural Information Processing Systems (NeurIPS).
  • Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang (2021). Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI).
  • Kraus, Maurice, Divo, Felix, Dhami, Devendra Singh, & Kersting, Kristian. (2025). xLSTM-Mixer: Multivariate time series forecasting by mixing via scalar memories. Thirty-ninth Conference on Neural Information Processing Systems (NeurIPS 2025).
Questions for authors

None that would materially affect my assessment.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.